The Scene
Juan M. Lavista Ferres is taking the work he does at Microsoft’s AI for Good Lab, where the technology is applied to solving real-world problems, into the public.
His first book, AI for Good, comes out Tuesday and is essentially a guide to using AI for complex issues, like facilitating humanitarian aid or assessing damage after an earthquake. Some of the chapter titles include “Monitoring Whales from Space” and “Social Networks of Giraffes.” The book, co-written by Microsoft’s Director of AI for Health William Weeks, is not a dry read about using Python.
Talent is pouring into the AI industry right now and every company is going to be impacted by fast, technological change. This book isn’t just useful for NGOs who want to solve global problems; It could also help companies that want to think outside the box about applying AI to problems.
Below is an edited transcript of an interview with Lavista Ferres, chief data scientist for the AI for Good Lab.
The View From Juan Lavista Ferres
Q: Microsoft has a lot of book authors — Kevin Scott, Brad Smith, the recently hired Mustafa Suleyman. Did you consult with any of them on this?
A: I talked to Brad. He told me that there’s a lot of work and that I had to be careful, because it was much more difficult than I thought it was. He was right. But he was fully supportive and actually wrote the foreward for the book.
Q: Did you use AI to help write it?
A: NO. As a non-native speaker, I do use AI to help me fix my grammatical errors all the time.
Q: Do you think there’s a shift happening where people are starting to see the positive use cases of AI?
A: People have been using AI for a long time, when they take an Uber or get a Netflix recommendation or do a Google search. They just didn’t call it AI. Since ChatGPT, people have been talking about AI a lot more. Like any new technology, people are afraid. People become less afraid when they start using the technology and start seeing the benefit. From that perspective, I think people are more optimistic.
Q: What are some ways people are using it that show the benefits?
A: Remember that 4 billion people in the world do not have access to doctors. That’s half of the world. Currently, the only solution we have right now as a society is to make sure that the doctors we have are more productive, so we can help get to more people.
There’s one initiative that we’re working on now in Mexico and Colombia focusing on retinopathy of prematurity, one of the leading causes of blindness in children. It affects babies who are born prematurely.
There are only 200,000 ophthalmologists in the world and millions of babies born prematurely. It’s physically impossible to diagnose the disease for every baby. We’ve got an AI model that runs on a phone that can diagnose the disease as well as an ophthalmologist.
Q: There are a lot of examples now of AI being able to do a better job than doctors on analyzing MRIs and other scans. But there’s some pushback on that technology, including from within the medical industry. Are you encountering resistance?
A: People are cautious. Of course we want these models to work perfectly and there is no such thing as a perfect model. We’re not doing this to replace the doctor. What we’re trying to do is help the doctor find those things that would be impossible for a medical doctor to find.
Q: What do you think of some of these deep tech applications of AI, such as AI and drug discovery?
A: It’s difficult to think of areas where AI cannot have an impact. We are working on subjects like protein folding and materials science. In material design, before it would take them years to run all these experiments and tell what is working and what’s not. Now you can use AI to say, of these 10,000 experiments, what are the 50 most likely to have an impact. Any area that is data-rich, AI can play a role.
Q: What about robotics? Do you think they can be used in AI for good?
A: Other than working with some robots with my kids with an Arduino, it’s not something we do at work. But once you have sensors, etc., at the end of the day, it’s a data problem. From an AI perspective, it’s not that different from other problems we’re working on.
It is a big opportunity, particularly in countries that will need that help. Particularly in developed countries, the ratio of people that are retirement age or older, robotics can help them in many ways. And then there’s the self-driving car category. My daughter is 10 years old. I still hope that she will have a self-driving car. The number one most dangerous thing that the majority of people in countries like the U.S. do — if they don’t smoke — is driving.
Q: The AI landscape is moving so fast. Do you think within a year, all of these examples in your book will look old?
A: Not so much. We have to remember that AI is not new. We’ve been working with some of these algorithms for the last 20 or 30 years. Even deep learning and neural nets have been around for over 30 years. What has dramatically changed is our ability to train these very big, large language models. When I started working in AI 20 years ago, I realized that natural language processing is even more difficult than working with images. Text is a really difficult problem. In large language models, it’s been a step function. If you had asked me 10 years ago, I would not have thought we’d get here this fast.
There are a lot of huge problems that we’re revisiting because before we couldn’t solve them, and now we can. But there’s still a lot of problems that haven’t changed. And we can still solve with things we were doing five years ago. So we’ll still run classification models in 15 years or 20 years.
Q: Since ChatGPT, the money and talent that is going into this has ramped up. When you look at all these problems around the world, does it excite you that there’s so much more talent that could work on them?
A: That definitely excites me. Even before ChatGPT, there was a huge demand for AI talent. Now that demand has exploded. And I think we’ll see even more of those papers. And as more people learn how to use this technology, they’ll work on solving problems that we couldn’t have even dreamed of solving.
Q: Five years from now, do you see superintelligent, general purpose models — some people might call it AGI -— helping to solve these problems around the world?
A: I usually try to stay away from the AGI conversation in general. I think these models will continue to become better. The reason I shy away from the AGI conversation is that the way people define AGI is through tests. One of the tests is the IKEA test. We used to have the Turing test. And of course a lot of LLMs will pass the Turing test. The IKEA test is when a [robotic AI] agent will go to your house, open a box and assemble furniture. I would not pass that test.
I focus on: We have this technology. It can be used to solve problems. The discussion should be about that.
These models like GPT are much more general now. We used to train very specific models. Now we have zero shot learning, where we don’t need a training set. Just put the information there and the models are able to solve problems that before, we would need to train models for. That’s already showing value from a general purpose perspective.
I’m still focusing on: We have a problem, we have a solution, we have a tool to solve a problem.
Models will continue to keep improving. Clearly what we saw in November 2022, it was a step function. But I don’t expect another big step function.
Q: Last question: What do you hope will be the impact of this book?
A: For researchers or people who are working on solving some of the world’s greatest challenges to use the book to first get inspired, but more importantly learn that some of these problems they have can be solved through AI. I want to try to help a much bigger set of organizations where they can read the book and say, “Hey, we could actually do something like this in Chapter 16.”